I never said AI would replace the musician. I never said it should. I never said it does.
I published a piece last week and the comments came in fast. Some agreed. A lot didn’t. A few pushed the argument somewhere I hadn’t taken it. That last group is the reason this piece exists, and I want to start by saying thank you, genuinely, to everyone who engaged, including and especially the ones who disagreed. This is exactly the conversation worth having, and you made it better than I did alone.
But there was one argument nobody made. Not the critics, not the allies, not the people who came in swinging and stayed to have the real conversation. And before we get to that, I want to correct something that ran through most of the pushback: I never argued that AI would, should, or does replace the musician. I never argued that the human steps out of the equation. What I argued was narrower and older than that. Every generation of production technology has been accused of killing music. Every time, the accusation was wrong. That was the premise. The comments largely debated a different one.
The argument nobody made is this: the demo tape was always the AI. And it changes everything about how we should be thinking about this.
I’ll come back to that. First, a few things I want to address from the thread directly.
Before I go further
Something a few people assumed that I want to address plainly: I am a musician. I have been one most of my life. I started on trombone, too small to reach sixth position for the low C, so I threw the slide and caught it with my foot. I also play bass, guitar, and noodle on piano. I write every lyric myself. I have made pop blues records, alt-country, dance, hip-hop, EDM, and alternative rock. I build each album as a somatic narrative, an emotional journey with a shape, not just a collection of songs sitting next to each other.
I also played in indie bands. In one, I brought the lyrics to a song and the guitarist wrote the music around them. We performed that song as part of a set we played more than 150 times in a single year. Part-time musicians with full-time day jobs, doing the thing we loved in whatever hours we could find. The songwriter and the musician were two different people on that stage, and the song was no less real for it.
I have also worked in AI professionally for three years. I say all of this not to establish credentials but because several arguments in the comments were aimed at someone who doesn’t make music. That’s not me. And it changes the lens.
The comments that got it right from the start
Several people made points I want to acknowledge before getting to the disagreements, because they deserve more than a like.
One commenter cut straight to the economic argument: “What’s actually killing music is the system built to cater to conglomerates and not pay artists. Majors are buying up catalogs rather than investing in new artists. Spotify’s pro rata system excludes streams under a certain number. AI has just exposed the fraud that this is.” That’s the thesis of the piece I published before this one, stated in three sentences.
Another noted that we’ve had versions of this conversation about notation diluting oral traditions, about whether instruments themselves were the first step in making music “artificial.” Someone else pointed out that atonal compositions, twelve-tone rows, and textural soundscapes were all declared “not real music” in their time. One person simply wrote: “Home taping is killing music.” Which is perfect.
One commenter brought up a detail I love: that people once argued using a rhyming dictionary was cheating, until Stephen Sondheim said he used one all the time, and suddenly it became acceptable. That’s the pattern in miniature. The tool is illegitimate until someone undeniable uses it. Then it’s a technique.
The techno producer who noted he uses the TR-808, TR-909, and TR-303 and builds everything himself made a fair point about soul and creative investment that I respect even where I disagree with the conclusion. The keyboardist from Austin who described AI as a killer tool for a solo songwriter who just needs to hear a full band on a demo, while flagging the messiness of prompting entire albums into existence and passing them off as original work, is describing a real tension I don’t dismiss.
To the BMI songwriter who raised model collapse and asked whether ASCAP and BMI could develop frameworks for AI attribution: that’s exactly the right institutional question, and I’ll come back to it.
The taste argument is a dead end
Several commenters made the case that AI music simply isn’t art: that art requires suffering, struggle, lived experience, and a nervous system on the other end. One put it plainly: “AI doesn’t experience or express anything, it just gathers and calculates an average of what looks like experience.”
I understand the feeling behind this. I don’t think it resolves anything.
Taste is subjective by definition. The person who cried at a film score generated with AI tools had a real emotional response. The person who felt nothing at a technically flawless human performance had a real absence of one. The music didn’t know the difference, and neither did the nervous system receiving it.
One commenter made the point more cleanly than I could: “If it entertains, inspires, or moves us emotionally, we’re not thinking in that moment, ‘I wonder what tools they used to create this.’” That’s not a dismissal of craft. It’s an honest description of how art actually lands.
The “is it real art” debate has never been settled and never will be, because it’s a values question disguised as an empirical one. The more interesting questions are the legal and economic ones underneath it. That’s where the thread got genuinely useful.
The mechanism argument: where I concede and where I hold
The sharpest challenge came in bullet points, and it deserves a direct response:
“The electric guitar didn’t scrape every recording ever made to learn how to replace the guitarist. The mechanism is not the same.”
This is correct, and I should have been more precise about where my historical analogy holds and where it breaks down. There is a real difference between a guitarist absorbing an influence over years of listening and practice, and a model ingesting millions of recordings in bulk without consent or compensation. Those are not identical processes and I won’t pretend they are.
The same commenter noted: “History repeating is an observation. It’s not an argument. Being wrong before doesn’t make you right now.” Also fair. The pattern of panic and eventual integration is real. It doesn’t guarantee the outcome this time. I accept that.
Here’s where I hold my ground. The underlying principle, that creativity is cumulative, that every artist is a distillation of everything they have heard and absorbed, is not new. Son House shaped Muddy Waters. Muddy Waters shaped Eric Clapton. Stevie Ray Vaughan studied Albert King so closely you can hear specific licks transposed almost note for note. Nobody asked permission. Nobody paid a licensing fee. The blues itself is a tradition built on communal borrowing across generations.
AI makes that process visible, compressed, and industrial in a way that makes people uncomfortable. What it doesn’t do is invent the concept. The question worth asking is not whether influence is legitimate. It always has been. The question is who gets paid when the distillation happens at scale.
The producer argument, and the best counter to it
Two commenters made versions of the same sharp point from different directions.
One stated it plainly: “Synths and drum machines changed what musicians play. Generative AI replaces the musician entirely. Those are not the same category.”
Another went further: “Quincy Jones didn’t just hand musicians a text prompt like ‘spooky 80s pop with funk bass’ and wait for a black box to spit out options. He directed with surgical, micro-level control using a deep vocabulary of harmony and rhythm. Producing is granular collaboration. Prompting an AI is macro-level delegation.”
The distinction between granular collaboration and macro-level delegation is real. Quincy Jones didn’t write a single song on Thriller and didn’t play a single note. He produced, arranged, and directed. The songs came from Michael Jackson and outside writers. The musicians executed. And nobody questions his creative authorship of that record. The producer role exists specifically because creative intent and physical execution are not the same thing.
But the commenter is right that a prompt is not the same as what Quincy brought to Westlake Recording Studios in 1982. I’m not making that claim. What I’m claiming is narrower: that execution and intent are separable, and that the intent in my process is mine. Whether that meets the bar of Quincy Jones is a different question entirely. It doesn’t have to, to be genuine.
The execution is different. The intention is not.
The most interesting argument in the thread
One commenter brought in the Yueji, a Confucian text from around 300 BC, which argued that music is the expression of the human heart in response to the world. He used it to draw a distinction that I think is the most precise version of the challenge to my original piece:
“AI as a tool, yes. Same line as everything on your list. But AI doing all of it, that’s somewhere the pattern hasn’t been before. That’s not the same argument. It’s a new one.”
The distinction he’s drawing is between augmentation and automation. Every technology I listed in the original piece augmented the human. The electric guitar amplified what Dylan’s hands were already doing. The 808 executed what Marvin Gaye’s instinct programmed into it. Auto-Tune processed a performance that existed first. At every point in that history, a human was inside the loop.
Another commenter made the same point more concisely: “Synths and drum machines changed what musicians play. Generative AI replaces the musician entirely. Those are not the same category.”
That’s the argument. And it’s the most honest version of the challenge to my original piece. It’s also where my personal practice becomes the only honest answer I can give. I don’t step out entirely. I am involved in every dimension of what gets made: tempo, mood, instrumentation, time signature, structure, recording approach, sonic signature. The decisions about what the piece is trying to say, what emotional arc it needs to move through, what serves the narrative and what gets cut, those are mine. Whether that constitutes genuine authorship is a question I can’t resolve for anyone else. I believe it does. I understand why others believe otherwise.
What I’d add: one commenter noted that so far AI music has largely produced imitations of existing pop songs rather than something genuinely new, and that this feels like a missed opportunity. That’s a fair observation about how most people are currently using the tool, not about what the tool is capable of. The same was said about samplers in the early 1980s, when most sample-based music was simply borrowing recognizable loops. What J Dilla and DJ Premier eventually did with the MPC was not that. The artists who push any tool beyond its obvious applications tend to arrive later. History suggests they will arrive here too. History doesn’t guarantee it.
A clarification I should have made from the start
Several commenters pushed back on the idea that AI replaces the musician. My premise was not that AI would, should, or does replace musicians. It was that the music industry treats every new production technology as a threat, argues that it is killing music, and is then proven wrong when the technology opens new genres and creative possibilities. The argument is about recorded and produced music, not about live performance. Those are different things and they deserve to be treated differently.
Here is the distinction that matters: a demo tape has always existed to get a song into the world so that a human could perform it. That separation between the songwriter, the producer, and the performer is not new. It is one of the oldest structures in popular music.
Willie Nelson wrote “Crazy” while struggling as a songwriter in Houston. He pitched it via a demo tape to Patsy Cline’s husband at Tootsie’s Orchid Lounge in Nashville. Cline wasn’t initially impressed, but she recorded it the next day and it became the most-played jukebox song in history. Nelson also wrote “Hello Walls,” “Funny How Time Slips Away,” and “Pretty Paper” before anyone knew his name as a performer. The songs existed first. The performances came later.
Kris Kristofferson worked as a janitor at Columbia Recording Studios in Nashville while writing the songs that would define his catalog, pitching demos whenever he could get someone to listen. “Me and Bobby McGee” was first recorded by Roger Miller in 1969. “Help Me Make It Through the Night” was made famous by Sammi Smith. “Sunday Mornin’ Comin’ Down” was recorded by Johnny Cash. In 1971, three of the five Grammy nominations for best country song were Kristofferson compositions, all made famous by performers who weren’t him. The demos that got them there were the tool that made it possible.
This is exactly what I use AI tools to do. I write the lyrics first. They have their own tempo and flow, their own internal rhythm. Then I work against the tool to find the right home for them: is this blues? Delta blues or Chicago blues? Is it country? 1950s swing or dance hall country? The genre isn’t a starting point, it’s a discovery. The lyrics lead and the production follows.
From there the decisions get granular. Tempo, mood, instrumentation, time signature, structure, sonic approach. I sit with a melody or a bridge for days. I try a sax solo and replace it with a cello. I ask whether this moment needs a full drum kit or just snare and kick. Every one of those decisions is made in service of what the song is trying to say, and none of them are made by the machine.
I don’t have the same time luxury I had in my twenties, when I could walk into a rehearsal space with a lyric and hand it to a guitarist and let the song find itself over weeks of practice. What I have now is a tool that gives me the band. The process is different. The intention behind the lyric is exactly the same.
That is not delegation. That is composition.
And the performance of that music by a human, live, in a room, is a completely separate conversation that I look forward to having in a future piece.
The real question: who gets paid
This is where the thread pushed me furthest, and where I want to spend the most time.
Sony has recently developed a system that can identify which recordings most influenced an AI-generated output and estimate the degree of contribution from each source. The method involves selectively “unlearning” a generated track from a model and measuring which training songs are most affected by that removal. In tests it achieved perfect identification of known training tracks. Sony Music and Universal Music Group have also partnered with SoundPatrol, a Stanford-affiliated research lab, to deploy neural fingerprinting tools capable of detecting the influence of original human-created recordings within AI-generated content.
This is significant, and it isn’t the only move the industry is making toward attribution infrastructure. Apple recently introduced AI Transparency Tags on Apple Music, a development I wrote about here [link], which signals that the streaming layer is beginning to build disclosure mechanisms even before the legal frameworks require them. The direction is clear. The details, and who benefits from them, are still being negotiated.
But here is where my position as a musician makes me more skeptical, not less.
If I write a song heavily influenced by Albert King without AI tools, no attribution technology flags it, no licensing fee is owed, no percentage is calculated. But the influence is just as real. The Sony technology isn’t solving the question of whether influence should be compensated. It’s only solving the question of whether AI-mediated influence can be detected. Those are different problems, and the debate is currently conflating them.
The legal question of whether AI-mediated influence is categorically different from human-mediated influence is genuinely unresolved. Suno’s own legal position draws exactly this distinction, arguing that a generated output that doesn’t sample actual sounds cannot infringe existing recordings as a matter of law, regardless of what it was trained on. That case is still working through the courts.
The technology exists to measure influence. The legal framework to define what that means is still being written. And the compensation system to actually get money to artists does not yet exist. I wrote about the longer history of that question in a piece published before this one, The Walled Garden and the Songwriter: AI, Labels, and the Long History of Getting Paid Last [link]. The short answer is that the history is not encouraging.
One commenter made the point that the BMI songwriter infrastructure emerged specifically because artists were incensed that radio stations played their music for free while making money from advertising. That’s exactly right, and it’s the model worth building on. The question is whether the institutions loudest in defending artists right now are building toward that outcome or positioning themselves to administer it on terms that favor the catalog holders over the creators. That question answers itself if you look at the history.
One more thing
A commenter who initially assumed I knew nothing about music, and then engaged more carefully after I explained my background, eventually landed somewhere I want to acknowledge directly:
“Maybe my actual concern is that instead of doing music the old way, we should now just become customers of a small group of infinitely powerful AI companies, run by a small group of anti-democratic, unimaginably greedy tech oligarchs that are literally destroying the planet.”
I don’t dismiss this. The concentration of power in a small number of AI companies is a real concern and one worth fighting over seriously. It’s also separate from whether the tools themselves can be used with artistic integrity. I can believe both that these tools can be part of a genuine creative practice and that the companies building them require serious scrutiny and accountability. Those positions are not in conflict. Holding only one of them is where the arguments in this thread, on both sides, sometimes went wrong.
The comment section pushed this piece somewhere better than the original. That doesn’t happen often enough. Thank you to everyone who showed up and engaged seriously, the allies, the skeptics, and the ones who came in swinging and stayed to have the real conversation. You know who you are and I’ll be tagging you below.
The conversation that actually matters, about compensation, attribution, consent, and who controls the infrastructure of music creation, is just beginning. If you haven’t read the Walled Garden piece yet, it’s linked below. If you read both, you’ll see that the history of technology in music and the history of who got paid from it are the same story told from two different angles.
The Walled Garden and the Songwriter: AI, Labels, and the Long History of Getting Paid Last